Important space and spectral feature self-learning lake COD inversion method

Abstract

The invention discloses a lake COD inversion method for important space and spectral feature self-learning, and the method comprises the steps: obtaining a hyperspectral image and a real COD value of a water sampling point, and generating an original training set and an original test set; constructing a COD parameter inversion model; adjusting the spectral data of the selected wavebands of the samples in the original training set, generating a reconstruction training set corresponding to the selected wavebands, training the COD parameter inversion model by using the original training set and the reconstruction training set corresponding to each selected waveband, selecting important wavebands, and obtaining the COD parameter inversion model based on the spectral data of the important wavebands of the samples in the original training set. Training the COD parameter inversion model to obtain a finally trained COD parameter inversion model; the most important wave band can be automatically selected, the most important spectral information can be acquired, the important spatial information can be automatically learned and screened, the influence of flares, aquatic plants and other impurities near sampling points on the spatial information is reduced, and the COD inversion precision is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing of lake water environment. Based on the hyperspectral images of lakes captured by an airborne hyperspectral camera of an unmanned aerial vehicle (UAV), a method for inverting the chemical oxygen demand (COD) of lakes with self-learning of important spatial and spectral features is provided, which is applicable to the inversion of the chemical oxygen demand (COD) in lakes. Background Art

[0002] The chemical oxygen demand (COD) is an important indicator for measuring the degree of organic pollution in water bodies. It is the amount of reducing substances that need to be oxidized in a water sample measured by a chemical method and is widely used in fields such as water treatment, environmental protection, and industrial wastewater monitoring. By regularly detecting the COD value, pollution sources can be discovered in time and corresponding treatment measures can be taken to ensure the sustainable utilization of water resources.

[0003] Compared with field sampling and laboratory analysis, satellite remote sensing images have a wide coverage range and are easily obtained, and are widely used in the field of water quality monitoring. However, affected by cloudy weather, satellite revisit cycles, and the spatial and spectral resolutions of the images, it is impossible to continuously obtain high-quality satellite remote sensing images of lakes with high spatial and high spectral resolutions. The emergence of UAV remote sensing technology provides a better solution. An airborne hyperspectral camera of a UAV can obtain high-spatial and high-spectral resolution lake water body images containing more texture and band information. However, the excessive number of bands in hyperspectral images will generate spectral information redundancy and reduce the inversion accuracy of COD parameters. Currently, many scholars use the Pearson coefficient or principal component analysis (PCA) to calculate the correlation between each band and the COD parameter value, and screen the important spectral band information in the hyperspectral image by sorting the correlations, and then use machine learning and other methods to establish an inversion model. However, these screening methods usually calculate the linear importance features between the band values and the COD values and cannot mine the non-linear importance features. Secondly, in terms of spatial information, while high spatial resolution brings refined features, it also causes information redundancy or noise. The current methods for obtaining the images corresponding to sampling points mostly use a grid of 5 5 (or 3 3) pixel points centered on the sampling point. However, sometimes there will be flare phenomena or interference from impurities such as waterweeds in some surrounding pixel points, reducing the usability of the images. At the same time, in the inversion method of water quality parameters, deep learning has shown good performance, can learn deeper non-linear features, and has strong generalization ability. In short, the existing research has not fully utilized the advantages of deep learning models for automatic and accurate screening and learning of important spatial and spectral features. Summary of the Invention

[0004] In view of the problems in the prior art that in the task of lake COD parameter inversion using hyperspectral images, the model does not adaptively screen important spatial and spectral features of the images, ignoring the relationship between important spatial and spectral features and COD values, and the poor accuracy of COD inversion using simple machine learning models, the present invention proposes a method (ISBLNet: Important Spatial and band learning Network) and device for lake COD inversion with self-learning of important spatial and spectral features, which can improve the inversion accuracy of COD in lakes and has better performance when compared with basic neural network models and random forest models.

[0005] The above technical problems of the present invention are mainly solved by the following technical solutions: A method for lake COD inversion with self-learning of important spatial and spectral features, comprising the following steps: Step 1: Obtain the hyperspectral image and the true COD value of the water sampling point, construct samples based on the hyperspectral image, use the true COD value as the label of the sample, and generate an original training set and an original test set; Step 2: Construct a COD parameter inversion model; Step 3: Train the COD parameter inversion model based on the original training set, calculate the coefficient of determination R² value corresponding to the COD parameter inversion model, adjust the spectral data of the selected bands of the samples in the original training set, generate a reconstructed training set corresponding to the selected bands, train the COD parameter inversion model using the reconstructed training sets corresponding to each selected band, calculate the coefficient of determination R² values corresponding to the COD parameter inversion models trained by each reconstructed training set, select several important bands that have the most important influence on the coefficient of determination R² value of the COD parameter inversion model, and train the COD parameter inversion model based on the spectral data of the important bands of the samples in the original training set to obtain the finally trained COD parameter inversion model; Step 4: Input the hyperspectral image to be predicted into the COD parameter inversion model to obtain the final predicted COD value.

[0006] Constructing samples based on the hyperspectral image as described above includes: performing radiometric correction on the hyperspectral image data, locating the water sampling point in the radiometrically corrected hyperspectral image, and cropping according to the set window range to obtain the cropped hyperspectral image, and using the cropped hyperspectral image as the sample.

[0007] The COD parameter inversion model as described above includes a spatial feature self-learning module, a convolutional module, and a fully connected module. The sample is input into the spatial feature self-learning module to obtain a feature map of the important spatial features after dimensionality reduction. The feature map of the important spatial features after dimensionality reduction is input into the convolutional module, and the convolutional module outputs a flattened feature vector; The flattened feature vector is input into the fully connected module, and the fully connected module outputs the predicted COD value.

[0008] The dimension of the sample input into the spatial feature self-learning module is batch channel height width. In the spatial feature self-learning module: Randomly extract the pixel points of the sample in the form of a window to generate window samples, and merge the window samples in the channel dimension according to the extraction order to obtain a window feature map. Perform a global max pooling operation on the window feature map to generate a pooled feature map. Perform a two-dimensional convolution operation on the pooled feature map to generate a spatial information weight factor with linear features. Perform a multiplication operation on the window feature map and the spatial information weight factor to generate an important spatial information feature map. Perform a two-dimensional convolution operation on the important spatial information feature map to restore the number of channels of the important spatial information feature map to the same as that of the sample input into the spatial feature self-learning module, and obtain the feature map of the important spatial features after dimensionality reduction.

[0009] As described above, the convolutional module of the COD parameter inversion model includes a first convolutional layer, a first ReLU activation function, a second convolutional layer, and a second ReLU activation function. The feature map of the important spatial features after dimensionality reduction is input into the first convolutional layer. The output features of the first convolutional layer are operated by the first ReLU activation function to further obtain the first non-linear features. The first non-linear features are input into the second convolutional layer. The output features of the second convolutional layer are operated by the second ReLU activation function to further obtain the second non-linear features. The height and width of the second non-linear features are flattened by the Flatten operation to obtain a flattened feature vector.

[0010] As described above, the fully connected module of the COD parameter inversion model includes a first fully connected layer, a third activation function, a second fully connected layer, a fourth activation function, and a third fully connected layer. The flattened feature vector is input into the first fully connected layer. The output features of the first fully connected layer are operated by the third activation function to introduce non-linear features and obtain the third non-linear features. The third non-linear features are input into the second fully connected layer. The output features of the second fully connected layer are operated by the fourth activation function to introduce non-linear features and obtain the fourth non-linear features. The fourth non-linear features are input into the third fully connected layer, and the third fully connected layer outputs the predicted COD value.

[0011] A computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the above inversion method are implemented.

[0012] A computer-readable storage medium stores a computer program thereon. When the computer program is executed by a processor, the steps of the above inversion method are implemented.

[0013] A computer program product includes a computer program. When the computer program is executed by a processor, the steps of the above inversion method are implemented.

[0014] The present invention has the following beneficial effects compared with the prior art: 1. Compared with using the Pearson coefficient to calculate the correlation between the reflectance values of each band in the hyperspectral image (each band is a spectral information) and the COD value, and using the PCA method to screen the important spectral features of the hyperspectral image, the COD parameter inversion model constructed by the neural network model in the present invention combines the band value replacement method, which can automatically select the most important bands, obtain the most important spectral information, avoid the limitations of manual selection, consider not only the linear relationship between the spectral information and the COD value, but also the non-linear relationship, make the screening of important spectral features more effective, and reduce the problem of spectral information redundancy in the hyperspectral image data.

[0015] 2. Compared with directly cropping the hyperspectral image within the range of 5 5 pixels near the sampling point, the present invention crops the image within the range of 7 7 near the sampling point, and performs multiple random samplings of the spatial range of the 7 7 window to generate multiple window samples of size 5 5. The respective window samples are merged in the channel dimension according to the extraction order to create a spatial attention mechanism, automatically learn and screen out important spatial information, reduce the influence of flares and impurities such as waterweeds near the sampling point on the spatial information, and improve the COD inversion accuracy.

[0016] 3. The present invention achieves higher accuracy compared with the basic network and the random forest method. Figure 1Figure 1 and Table 1 show the fitting effect diagrams of the present invention on the original test set (compared with the basic network, the spatial feature self-learning module and the spectral feature self-learning module are removed). It can be seen that after adding the spatial feature self-learning module and the spectral feature self-learning module, the present invention has an improvement of 0.026 in the coefficient of determination R 2 compared with the basic network, and the RMSE is reduced by 0.2628 mg / L. Compared with the random forest model, R 2 is increased by 0.0854, and the RMSE is reduced by 0.8228 mg / L. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Figure 2 shows the fitting result diagrams of the coefficient of determination and the root mean square error of the present invention, the basic network, and the random forest on the original test data set (R 2 is the coefficient of determination, and RMSE is the root mean square error). Among them, (a) is the fitting result diagram of the coefficient of determination and the root mean square error of the present invention on the original test data set; (b) is the fitting result diagram of the coefficient of determination and the root mean square error of the basic network on the original test data set; (c) is the fitting result diagram of the coefficient of determination and the root mean square error of the random forest on the original test data set; Figure 2 Figure 3 is the flow chart of the present invention; Figure 3 Figure 4 is the noise schematic diagram of the water sampling point; among them, (a) is the noise schematic diagram of the water surface of the water sampling point with organisms such as waterweeds; (b) is the noise schematic diagram of the water surface of the water sampling point with flare; Figure 4 Figure 5 is the structural schematic diagram of the spatial feature self-learning module; Figure 5 Figure 6 is the structural schematic diagram of the convolutional module; Figure 6 Figure 7 is the structural schematic diagram of the fully connected module. DETAILED DESCRIPTION OF THE INVENTION

[0018] In order to facilitate the understanding and implementation of the present invention by those of ordinary skill in the art, the present invention will be further described in detail below with reference to the embodiments. It should be understood that the embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0019] Embodiment 1

[0020] An important lake COD inversion method for spatial and spectral feature self-learning (ISBLNet) includes the following steps.

[0021] Step 1: Obtain the hyperspectral image and the true COD value of the water sampling point, construct samples based on the hyperspectral image, use the true COD value as the label of the sample, and generate the original training set and the original test set.

[0022] This step mainly includes the acquisition of water samples and hyperspectral images, the acquisition of true COD values through laboratory COD measurement, water body image processing, and dataset division, specifically including the following steps: The acquisition of hyperspectral image data is carried out through the following steps: Use a water sampler to obtain water samples at each water sampling point in the target water area. At the same time, use a drone equipped with a hyperspectral camera to image the target water area to obtain hyperspectral image data. Perform radiometric calibration on the hyperspectral image data. Locate the water sampling points in the radiometrically calibrated hyperspectral image and crop them according to the set window range to obtain the cropped hyperspectral image.

[0023] The acquisition of true COD values is carried out through the following steps: Send the water samples obtained at each water sampling point to the laboratory to measure their chemical oxygen demand (COD) values as the corresponding true COD values.

[0024] Take the cropped hyperspectral image as the sample and the corresponding true COD value as the label. Randomly divide the samples to construct the original training set and the original test set. In this embodiment, the numbers of the original training set and the original test set are 76 and 32 respectively.

[0025] Step 2: Construct a COD parameter inversion model. The COD parameter inversion model includes a spatial feature self-learning module, a convolutional module, and a fully connected module, as Figures 4 to 6 .

[0026] Step 2.1: Input the sample into the spatial feature self-learning module to obtain a feature map of the important spatial features after dimensionality reduction.

[0027] Step 2.1.1: Input the sample into the spatial feature self-learning module batch by batch. The dimension of the sample is batch_size channels bands height width, and the dimension size is 8 125 7 7. The number of batches refers to the number of samples input into the spatial feature self-learning module each time during training. The initial number of channels is the number of bands of the sample.

[0028] Step 2.1.2: Randomly extract pixel points of each sample in a batch (8 in the present invention) in the form of a window to generate n channels height width of 125 5 5 window samples (the dimension size of each window is 125 5 5. In the present invention, n is 441. The window samples are combined in the channel dimension according to the extraction order. Finally, a batch will generate a window feature map with a dimension of 8 125n 5 5.

[0029] Step 2.1.3: Perform global max pooling operation on the window feature map to generate a pooling feature map with a dimension of 8 125n 1 1.

[0030] Step 2.1.4: Perform a two-dimensional convolution operation with a convolution kernel size of 1 1 on the pooling feature map to generate a spatial information weight factor with a dimension of 8 125n 1 1 with linear features.

[0031] Step 2.1.5: Perform a multiplication operation on the window feature map and the spatial information weight factor to generate an important spatial information feature map with a dimension of 8 125n 5 5.

[0032] Step 2.1.6: Perform a two-dimensional convolution operation with a convolution kernel size of 1 1 on the important spatial information feature map to reduce the dimension of the important spatial information feature map and restore the number of channels of the important spatial information feature map to the number of channels of the sample input to the spatial feature self-learning module (i.e., restore its number of bands to the original 125), generating a feature map of the reduced-dimensional important spatial features with a dimension of 8 125 5 5.

[0033] As shown in (a) of Figure 3 , there are organisms such as aquatic plants on some water surfaces in the figure, and due to the problem of the incident angle of sunlight, there will also be some flare phenomena, as shown in Figure 3As shown in (b), these phenomena cause the reflectivity of this pixel to change and no longer be the reflectivity of water, resulting in a decrease in the inversion accuracy of the COD parameter. Usually, these areas are actively avoided when cropping the image near the sampling point. However, sometimes the pixels of these phenomena are discontinuous and cannot be completely avoided, which will cause some irrelevant spatial information to be mixed into the cropped hyperspectral image. Randomly extracting the image near the sampling point using a window can randomly generate multiple window samples with different spatial position combinations. Then, using the spatial self-learning feature module to learn the important spatial information can make the neural network ignore some irrelevant information and improve the inversion accuracy of the COD parameter.

[0034] Step 2.2: The feature map of the important spatial features after dimensionality reduction is input into the convolution module, and the convolution module outputs a flattened feature vector. The convolution module includes a first convolutional layer, a first ReLU activation function, a second convolutional layer, and a second ReLU activation function.

[0035] Step 2.2.1: The feature map of the important spatial features after dimensionality reduction is input into the first convolutional layer. After the operation of the first convolutional layer (the number of input channels is 125, the number of output channels is 64, and the convolutional kernel size is 3 3, and the padding is 0), the dimension size of the output feature of the first convolutional layer is 8 64 3 3.

[0036] Step 2.2.2: The output feature of the first convolutional layer undergoes the operation of the first ReLU activation function to further obtain the first non-linear feature. The dimension size of the first non-linear feature is 8 64 3 3.

[0037] Step 2.2.3: The first non-linear feature is input into the second convolutional layer. After the operation of the second convolutional layer (the number of input channels is 64, the number of output channels is 32, and the convolutional kernel size is 3 3, and the padding is 0), the dimension size of the output feature of the second convolutional layer is 8 32 1 1.

[0038] Step 2.2.4: The output feature of the second convolutional layer undergoes the operation of the second ReLU activation function to further obtain the second non-linear feature. The dimension size of the second non-linear feature is 8 32 1 1.

[0039] Step 2.2.5: The height and width of the second non-linear feature are flattened through the Flatten operation, and the output of the second non-linear feature is flattened into a one-dimensional vector, obtaining a flattened feature vector with a size of 8 32 (8 is the batch size), which is convenient for connecting to the fully connected layer.

[0040] Step 2.3: Input the flattened feature vector into the fully connected module. The fully connected module outputs the predicted COD value. The fully connected module includes a first fully connected layer, a third activation function, a second fully connected layer, a fourth activation function, and a third fully connected layer.

[0041] Step 2.3.1: The flattened feature vector is input into the first fully connected layer. After the operation of the first fully connected layer, the dimension size changes from 8 32 to 8 64 to learn high-level features.

[0042] Step 2.3.2: The output features of the first fully connected layer are operated by the third activation function to introduce non-linear features and further obtain the third non-linear feature. The dimension size of the third non-linear feature is 8 64.

[0043] Step 2.3.3: The third non-linear feature is input into the second fully connected layer. After the operation of the second fully connected layer, the feature size changes from 8 64 to 8 128 to further learn high-level features.

[0044] Step 2.3.4: The output features of the second fully connected layer are operated by the fourth activation function to introduce non-linear features and further obtain the fourth non-linear feature. The dimension size of the fourth non-linear feature is 8 128.

[0045] Step 2.3.5: The fourth non-linear feature is input into the third fully connected layer. After the operation of the third fully connected layer, the feature size changes from 8 128 to 8 1, mapping the feature to an output predicted COD value.

[0046] Step 3: In the spectral feature self-learning module, train the COD parameter inversion model based on the original training set, calculate the coefficient of determination R² value corresponding to the COD parameter inversion model, adjust the spectral data of the selected bands of the samples in the original training set to generate a reconstructed training set corresponding to the selected bands, train the COD parameter inversion model using the reconstructed training sets corresponding to each selected band, calculate the coefficient of determination R² value corresponding to the COD parameter inversion model trained by each reconstructed training set, select several important bands that have the most significant impact on the coefficient of determination R² value of the COD parameter inversion model, and train the COD parameter inversion model based on the spectral data of the important bands of the samples in the original training set to obtain the finally trained COD parameter inversion model.

[0047] The training of the COD parameter inversion model is to train the COD parameter inversion model based on minimizing the loss function: use the mean squared error (MSE) loss function to calculate the loss between the predicted COD value and the true COD value, and the loss is backpropagated for the optimization and update of the parameters of the COD parameter inversion model. During training, the loss will gradually decrease until convergence (the loss value is stable). The number of epochs for network training in the present invention is 50, and finally, an accurate predicted COD value is output.

[0048] Step 3.1: Train the COD parameter inversion model using the original training set, test the trained COD parameter inversion model using the original test set, and calculate the coefficient of determination R² value of the COD parameter inversion model.

[0049] Step 3.2: Randomly shuffle the spectral information of the selected bands of the samples in the original training set to regenerate a reconstructed training set, further obtain the reconstructed training sets corresponding to each selected band, train the COD parameter inversion model using the reconstructed training sets corresponding to each selected band respectively, test the COD parameter inversion models trained by each reconstructed training set using the original test set respectively, and calculate the coefficient of determination R² value of the COD parameter inversion models trained by each reconstructed training set.

[0050] Step 3.3: Select the first set number of selected bands that have a greater impact on the coefficient of determination R² value (the first 20 bands with a greater impact on the coefficient of determination R² value in this embodiment) as important bands, and train the COD parameter inversion model based on the spectral data of the important bands of the samples in the original training set to obtain the finally trained COD parameter inversion model.

[0051] Step 4: Input the hyperspectral image to be predicted into the COD parameter inversion model to obtain the final predicted COD value.

[0052] The finally trained COD parameter inversion model, basic network, random forest and other models are tested using the original test set, and the predicted COD values are output for accuracy evaluation and comparative analysis. The results of the comparative analysis are shown in Table 1. It can be seen from the table that the present invention is superior to the basic network and random forest models in terms of the coefficient of determination R² and the root mean square error RMSE.

[0053] Table 1 Results of testing each model based on the original test set

[0054] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0055] Embodiment 2 In this embodiment, a computer device is further provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the above method embodiments are implemented.

[0056] Embodiment 3 In this embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0057] Embodiment 4 In this embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0058] It should be noted that the embodiments described in the present invention are only examples of the spirit of the present invention. Those skilled in the art to which the present invention pertains can make various modifications or supplements to the described embodiments or use similar ways to replace them, but will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. An inversion method for lake COD with self-learning of important spatial and spectral features, characterized in that It includes the following steps: Step 1: Obtain the hyperspectral image and the true COD value of the water sampling point, construct samples based on the hyperspectral image, use the true COD value as the label of the samples, and generate the original training set and the original test set; Step 2: Construct a COD parameter inversion model; Step 3: Train the COD parameter inversion model based on the original training set, calculate the coefficient of determination R² value corresponding to the COD parameter inversion model, adjust the spectral data of the selected bands of the samples in the original training set, generate the reconstructed training set corresponding to the selected bands, use the reconstructed training sets corresponding to each selected band to train the COD parameter inversion model, calculate the coefficient of determination R² values corresponding to the COD parameter inversion models trained by each reconstructed training set, select several important bands that have the most important influence on the coefficient of determination R² value of the COD parameter inversion model, and train the COD parameter inversion model based on the spectral data of the important bands of the samples in the original training set to obtain the finally trained COD parameter inversion model; Step 4: Input the hyperspectral image to be predicted into the COD parameter inversion model to obtain the final predicted COD value.

2. The method for inverting lake COD with self-learning of important spatial and spectral features according to claim 1, characterized in that, The construction of samples based on the hyperspectral image includes: performing radiometric correction on the hyperspectral image data, locating the water sampling point in the radiometrically corrected hyperspectral image, and cropping according to the set window range to obtain the cropped hyperspectral image, and using the cropped hyperspectral image as a sample.

3. The lake COD inversion method with self-learning of important spatial and spectral features according to claim 1, characterized in that, The COD parameter inversion model includes a spatial feature self-learning module, a convolutional module, and a fully connected module. The sample is input into the spatial feature self-learning module to obtain a feature map of the important spatial features after dimensionality reduction; The feature map of the important spatial features after dimensionality reduction is input into the convolutional module, and the convolutional module outputs a flattened feature vector; The flattened feature vector is input into the fully connected module, and the fully connected module outputs the predicted COD value.

4. The method for inverting lake COD with self-learning of important spatial and spectral features according to claim 3, wherein, The dimension of the samples input into the spatial feature self-learning module is batch channels height width. In the spatial feature self-learning module: Randomly extract the pixel points of the sample in the form of a window to generate window samples, and merge the window samples in the channel dimension according to the extraction order to obtain a window feature map. Perform a global max pooling operation on the window feature map to generate a pooled feature map. Perform a two-dimensional convolution operation on the pooled feature map to generate a spatial information weight factor with linear features. Perform a multiplication operation on the window feature map and the spatial information weight factor to generate an important spatial information feature map. Perform a two-dimensional convolution operation on the important spatial information feature map to restore the number of channels of the important spatial information feature map to the same as that of the sample input into the spatial feature self-learning module, and obtain a feature map of the important spatial features after dimensionality reduction.

5. The lake COD inversion method with self-learning of important spatial and spectral features according to claim 3, characterized in that, The convolutional module of the COD parameter inversion model includes a first convolutional layer, a first ReLU activation function, a second convolutional layer, and a second ReLU activation function. The feature map of the important spatial features after dimensionality reduction is input into the first convolutional layer. The output feature of the first convolutional layer undergoes a first ReLU activation function operation to further obtain a first non-linear feature. The first non-linear feature is input into the second convolutional layer. The output feature of the second convolutional layer undergoes a second ReLU activation function operation to further obtain a second non-linear feature. The height and width of the second non-linear feature are flattened through a Flatten operation to obtain a flattened feature vector.

6. The method for inverting lake COD with self-learning of important spatial and spectral features according to claim 3, wherein The fully-connected module of the COD parameter inversion model includes a first fully-connected layer, a third activation function, a second fully-connected layer, a fourth activation function, and a third fully-connected layer. The flattened feature vector is input into the first fully-connected layer. The output feature of the first fully-connected layer undergoes an operation of the third activation function to introduce non-linear features and obtain a third non-linear feature. The third non-linear feature is input into the second fully-connected layer. The output feature of the second fully-connected layer undergoes an operation of the fourth activation function to introduce non-linear features and obtain a fourth non-linear feature. The fourth non-linear feature is input into the third fully-connected layer, and the third fully-connected layer outputs a predicted COD value.

7. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the inversion method described in any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the inversion method described in any one of claims 1 to 6.

9. A computer program product comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the inversion method described in any one of claims 1 to 6.

Images